A Gloss-driven Indian Sign Language Production System Using Learned Pose Representations

1 Introduction

Sign languages (SLs) serve as the primary means of communication for Deaf and hard-of-hearing communities. According to Ethnologue, Indo-Pakistani SL is the most widely used SL globally 1 . In India alone, an estimated 63 million individuals experience hearing impairments 2 , yet access to inclusive education and public services remains severely limited. While SL interpretation has traditionally relied on human interpreters, their limited availability, high cost, and practical constraints highlight the urgent need for automatic SLP systems. By enabling real-time translation of spoken or written content into SL, SLP systems can democratize access to information and directly support the United Nations Sustainable Development Goals (SDGs) 3 , particularly SDG 4 (Quality Education), SDG 8 (Decent Work), and SDG 10 (Reduced Inequalities) [Fink et al., 2023]. This vision aligns with India's National Education Policy (NEP) 2020 4 , which emphasizes accessibility and equity in learning and public service delivery.

1 https://www.ethnologue.com/guides/ethnologue200

2 https://www.who.int/india/campaigns/world-hearing-day-2023

To the best of our knowledge, only a limited number of [Sugandhi

et al., 2020; Kumar et al., 2021]. These systems primarily rely on HamNoSys [Hanke, 2004], which requires expert linguistic annotation and thus limits scalability. Moreover, they typically generate utterances by concatenating word-level (isolated) signs without modeling co-articulation 5 , resulting in temporally discontinuous and visually unnatural avatar animations. Recent deep learning based approaches [Stoll et al., 2018; Zelinka and Kanis, 2020; Saunders et al., 2020; Walsh et al., 2024] have demonstrated promising results by leveraging large, well-annotated datasets such as RWTHPHOENIXWeather-2014T (PHOENIX14T) [Camgoz et al., 2018] and others [Duarte et al., 2021; Zhou et al., 2021; Albanie et al., 2021]. However, comparable large-scale, wellannotated continuous datasets are unavailable for ISL. As a result, ISL remains a severely low-resource language for SLP. Additionally, many existing systems require retraining to incorporate vocabulary expansion, further limiting scalability and real-world deployment.

3 https://sdgs.un.org/goals

4 https://www.education.gov.in/en/nep/about-nep

5 Smooth transitions between isolated signs.

Based on the output modality, SLP methods can be broadly categorized into three groups: (1) video-based (generate RGB frame sequences) [Saunders et al., 2022; Shi et al., 2024](2) keypoint-based (generate sequences of human pose keypoints) [Zelinka and Kanis, 2020; Saunders et al., 2020; Xie et al., 2024; Shi et al., 2024; Tang et al., 2025c; Wang et al., 2025; Tang et al., 2025b]; and (3) avatar-based (produce animations of digital 3D humanoid avatars) [Zuo et al., 2024; Yin et al., 2024; Sugandhi et al., 2020]. Across paradigms, generating temporally coherent pose keypoint sequences from glosses 6 remains a core challenge, as pose generation serves as a common intermediate step.

Sign language is inherently visual and relies not only on isolated signs but also on smooth transitions between them (co-articulation). Proper co-articulation is therefore essential for synthesizing meaningful continuous signs [Tang et al., 2025a]. In low-resource ISL settings, well-annotated continuous datasets are scarce, whereas several isolated sign datasets [Sridhar et al., 2020; Joshi et al., 2022; Patra et al., 2025b] are publicly available. In this work, we leverage these isolated resources to construct sentence-level signs.

Gloss-free end-to-end models [Saunders et al., 2020; Tang et al., 2025b] that directly regress pose sequences often suffer from temporal inconsistency and regression-to-the-mean effects. Discretizing pose space using Vector Quantized Variational Autoencoder (VQ-VAE) [Van Den Oord et al., 2017] alleviates these issues by learning discrete pose tokens [Rastgoo et al., 2024]. However, subtle hand and finger motions, critical for SL, are frequently attenuated due to their highfrequency and low-amplitude nature [Zhang et al., 2025]. To address this, we disentangle body, left-hand, and right-hand keypoints and normalize hand poses to better capture handshape and fine-grained finger articulation.

Recent dictionary-based approaches [Walsh et al., 2024; Zuo et al., 2024] demonstrate that stitching isolated signs avoids regression-to-the-mean issues and enables scalable vocabulary expansion. Inspired by this paradigm, a dictionarybased SLP system is proposed that produces natural continuous signing with proper co-articulation. This work makes the following key contributions:

• A modular and scalable SLP system for low-resource sign languages, demonstrated on Indian Sign Language, that generates sentence-level continuous signs without requiring large continuous datasets.
• Sign-Pose-VQ-VAE: a VQ-VAE-based model that learns discrete body, left-hand, and right-hand pose codebooks, enabling fine-grained hand articulation and natural co-articulation.

6 A written label representing a unit sign for annotation.

• A simple baseline method to transfer pose trajectories to 3D avatar animation through bone-angle replication for enhanced end-user understanding.
• Comprehensive evaluation demonstrating strong performance, including a gloss BLEU-4 of 47.20 on ISL, ISLRTC 7 certified interpreter-rated quality of 4.33/5, and state-of-the-art results among keypoint-based approaches on PHOENIX14T (BLEU-4 score of 10.03).

This project is part of a larger initiative to enable bidirectional translation between spoken language and ISL in educational settings, initiated by the ISL unit of FDMSE, RKMVERI, Coimbatore, a teaching institution for disability management in India. In addition to ISL domain experts, certified ISL interpreters (ISLRTC-certified) and professional linguistic experts contributed by annotating gloss translations from English sentences and evaluating the system.

The rest of the paper is organised as follows: Section 2 discusses the related works, and Section 3 describes the proposed method. Experimental evaluation is presented in Section 4, and the paper concludes in Section 5.

2 RelatedWorks

Prior SLP methods [Hu et al., 2022; Cox et al., 2002; Glauert et al., 2006] which rely on motion capture data requires specialized, expensive equipments, restricting implementations in low-resource settings. Other methods [Zwitserlood et al., 2004; Goyal and Goyal, 2016; Sugandhi et al., 2020; Kumar et al., 2021] use the Hamburg Notation System (HamNoSys) [Hanke, 2004], which often produces robotic signing with poor co-articulation, resulting in unnatural output that is generally not well received by the deaf community.

With deep learning advances, researchers are exploring its potential in SLP. The neural machine translation based approach [Stoll et al., 2018] splits the task into Text-toGloss (T2G) using a Recurrent Neural Network, Gloss-toPose (G2P) using a gloss-2D pose dictionary, and Pose-toSign using a Generative Adversarial Network (GAN) [Goodfellow et al., 2014]. However, it ignores hand keypoints, crucial in SL. Progressive Transformers [Saunders et al., 2020] uses two Transformers [Vaswani et al., 2017] for T2G and G2P. Tang et al.'s [Tang et al., 2022] GEN-OBT enhances a transformer gloss encoder and applies a reverse-translation objective to better align gloss and pose representations. Viegas et al. [Viegas et al., 2023] incorporate facial expressions as non-manual features in a transformer-based SLP model.

Recent works in SLP have moved rapidly from glossconditioned encoder-decoder pipelines toward diffusionbased generative models. Xie et al. [Xie et al., 2024] used Pose-VQVAE and a discrete denoising diffusion model (G2PDDM), using vector-quantized codes and a CodeUNet to capture spatial-temporal structure while handling variable output length. GCDM [Tang et al., 2025c], and Sign-IDD [Tang et al., 2025b], are other notable works that show the effectiveness of diffusion mechanisms for the SLP task. Despite their promising results, these models often generate temporally inconsistent pose sequences. In contrast, we select a convolutional architecture along the temporal dimension for the SignPose-VQ-VAE, which effectively captures local temporal dependencies and ensures smoother, more temporally consistent pose transitions throughout the generated sign sequences.

7 https://islrtc.nic.in/directory-of-isl-interpreters/

VQ-VAE [Van Den Oord et al., 2017] based approach T2SGPT [Yin et al., 2024] uses a dynamic VQ-VAE to learn variable-length codes and use GPT networks to predict codes and lengths, from which they generate the continuous sign sequence. Here, they considered body keypoints, including hands, whereas we decoupled the keypoints into multiple streams, which leads to a wide variety of encoded poses in a relatively small codebook. Recently, Zuo et. al. [Zuo et al., 2024] proposed a 3D avatar based approach. They segmented the continuous SL videos and created a gloss-3D avatar sign dictionary. They concatenate these 3D avatar signs using linear interpolation method. Similarly, in ISL, Patra et al. [Patra et al., 2025a] leveraged an isolated keypoint-based sign dictionary and linear interpolation for co-articulation to generate continuous sign sequences. In contrast to these dictionarybased works, our approach only stores the codebook indices (tokens) in the dictionary, which is much more efficient in real-world applications.

3 ProposedMethodology

The proposed SLP system is formulated as follows: given a spoken-language sentence X = [x1, x2, . . . , xW] consisting of W words, the goal is to generate an output sequence Y = [y1, y2, . . . , yO] of length O, where each yi represents a SL output unit, such as a skeleton pose, a photorealistic frame, or, in our case, an avatar animation frame. The proposed system consists of four stages: gloss translation, gloss-code dictionary generation, co-articulated sign generation, and avatar animation generation. An overview of the proposed system is shown in Figure 1, where English text is translated into a sequence of glosses. The corresponding codes are then retrieved, concatenated, and decoded to generate a co-articulated sign sequence, which is finally converted into a signing avatar animation.

3.1 Gloss Translation

The goal of this stage is to translate spoken-language text into a sequence of sign glosses. Due to the lack of a wellannotated text-gloss parallel corpus for ISL, neural machine translation approaches are not directly applicable. Therefore, a rule-based method is adopted, and a large language model

(LLM)-based approach is additionally explored.

Rule-based: Each spoken-language sentence is parsed into a syntactic tree using the Stanza library 8 with part-of-speech tagging. The sentence is then reordered into a subject-objectverb (SOV) structure, followed by the application of ISL grammatical rules inspired by Sugandhi et al. [Sugandhi et al., 2020] to generate ISL gloss sequences.

LLM-based: Inspired by Inan et al. [Inan et al., 2024], an LLM-based approach for text-to-ISL-gloss translation is employed. With the assistance of ISL interpreters, a small corpus of 300 English sentences with their corresponding ISL gloss translations is manually created. The LLM is then prompted to translate unseen sentences into ISL glosses based on these examples.

3.2 Gloss-Code Dictionary Creation

Dictionary-based approaches [Saunders et al., 2022; Walsh et al., 2024; Zuo et al., 2024; Inan et al., 2024] typically store high-dimensional visual representations (e.g., frames, videos, animations, or pose sequences), which are memory-intensive and limit scalability. Instead, Sign-Pose-VQ-VAE is proposed, which learns compact, discrete representations of body poses and handshapes by disentangling body, left-hand, and right-hand keypoints into separate codebooks. Only the resulting codebook indices (tokens) are stored in a gloss-code dictionary, while continuous pose sequences with smooth coarticulation are reconstructed by a convolutional decoder. An overview of this stage is illustrated in Figure 2.

Sign Pose Vector Quantized Variational Autoencoder (Sign-Pose-VQ-VAE): Given a pose sequence S = [s1, . . . , sF ] extracted from a sign video, where si ∈ RK×3 denotes K 3D keypoints, S is decoupled into three streams: body, left hand, and right hand. To preserve hand-body coordination, wrist joints and selected 2 metacarpophalangeal (MCP) joints from each hand are included in the body stream, resulting in 15 body keypoints. These three additional keypoints per hand are retained to later reattach the hands to the body. Hand keypoints are normalized per frame to achieve invariance to position, rotation, and scale.

The resulting streams are Sb ∈ RF×15×3 , Sl ∈ RF×21×3 , and Sr ∈ RF×21×3 . Each stream is processed by an independent VQ-VAE with encoder Ej and decoder Dj (j ∈ {b, l, r}). The encoder consists of three 1D convolutions followed by a residual block [He et al., 2016] and ReLU activations, producing latent features Z j = [zj1 , . . . , z jF/2 ], where zj i ∈ Rdc . Latents are quantized using a learnable codebook C j = {cjn }Nn=1 :

Codebooks are updated using exponential moving average (EMA) [Razavi et al., 2019] with codebook reset [Williams et al., 2020] to prevent collapse. The decoder Dj reconstructs pose sequences using residual blocks, nearest-neighbor upsampling, four 1D convolution layers, and ReLU activations.

8 https://stanfordnlp.github.io/stanza/

Reconstruction is optimized with a smooth L1 loss:

Since the EMA variant is used, the embedding loss is omitted, while the commitment loss is retained:

where sg[·] denotes the stop-gradient operator and β is the weighting coefficient for the commitment loss.

After combining all three streams, the final objective function of Sign-Pose-VQ-VAE is:

The proposed Sign-Pose-VQ-VAE is trained on randomly sampled pose segments from the isolated ISL dataset FDMSE-ISL [Patra et al., 2025b]. After training, each isolated gloss v is encoded into three discrete index sequences (I b , I l , I r ), where I j = [ij1 , . . . , ijF/2 ] and ijm ∈ {1, . . . ,N}. The resulting gloss-code dictionary maps each gloss to its corresponding discrete pose tokens (Figure 2).

3.3 Co-Articulated Sign Generation

During training, the Sign-Pose-VQ-VAE decoders are trained to learn smooth pose and handshape transitions; when applied jointly, they implicitly model co-articulation across consecutive signs. An overview of this module is shown in Figure 2. Given a gloss sequence V = [v1, . . . , vG], the corresponding discrete code sequences are retrieved from the gloss-code dictionary. For each gloss vi, three code sequences ( ˆ Z b i , ˆ Z l i , ˆ Z r i ) are obtained. These sequences are concatenated independently for the body and both hands, ˆ Z ′ = [||Gi=1 ˆ Z b i , ||Gi=1 ˆ Z l i , ||Gi=1 ˆ Z r i ], where || denotes temporal concatenation. Each stream of ˆ Z ′ is decoded by its corresponding VQ-VAE decoder to generate continuous body pose and handshape sequences. Finally, the handshape streams are coupled with the body pose stream by aligning the shared keypoints, yielding a temporally smooth, co-articulated sign sequence.

3.4 Avatar Animation Generation

Given a continuous pose sequence, the goal is to generate a corresponding avatar animation. Since fine-grained hand articulation is critical for SL, the upper body, particularly the arms and hands, is animated. A generic humanoid avatar from Mixamo 9 is utilized, and human joint rotation constraints (e.g., three degrees of freedom for shoulders and one for elbows) are enforced to compute per-bone Euler angles. These angles are then applied to the avatar skeleton to produce the final animation.

4 Experiments &Results

4.1 Dataset

The proposed system is evaluated on two datasets: the FDMSE-ISL dataset [Patra et al., 2025b] and the RWTHPHOENIXWeather-2014T (PHOENIX14T) dataset [Camgoz et al., 2018]. The FDMSE-ISL dataset [Patra et al., 2025b] contains 40,033 isolated ISL videos spanning 2,002 unique glosses. This dataset is used to train Sign-Pose-VQ-VAE and to construct the gloss-code dictionary. To evaluate sentencelevel sign generation, a Synthetic Indian Continuous Signs dataset (SynICS-20k) is constructed by concatenating isolated signs corresponding to 20,000 generic English subtitle sentences derived from Hindi movies from OpenSubtitles 10 . Linear interpolation is applied at sign boundaries. This dataset is referred to as SynICS-20k and is split into training (70%), validation (10%), and test (20%) sets.

9 https://www.mixamo.com

The RWTH-PHOENIXWeather-2014T (PHOENIX14T) dataset [Camgoz et al., 2018] consists of 8,257 German sentences with corresponding gloss annotations and sign videos over 1,115 unique glosses, and is used for state-of-the-art benchmarking. This dataset has no isolated gloss-level video annotations and is segmented using the CTC-based force alignment method proposed by Zuo et al. [Zuo et al., 2024].

4.2 Experiment Settings

Sign-Pose-VQ-VAE is implemented in PyTorch and trained using the AdamW optimiser with an initial learning rate of 2 × 10-4 , cosine annealing, 20 warm-up epochs, early stopping, and commitment loss coefficient β = 0.02. The model is trained for 7.5k epochs on SynICS-20k and 30k epochs on PHOENIX14T. Each stream is configured with a codebook of size N = 1536 and latent dimensionality dc = 128, selected via ablation. All experiments are conducted on an Intel i913900K CPU with an NVIDIA RTX 4090 GPU. MediaPipe Holistic 11 is used for 3D keypoint estimation on the FDMSEISL dataset. For fair comparison on the PHOENIX14T dataset, HRNet [Sun et al., 2019] is employed for 2D pose estimation, and a Recurrent Neural Network (RNN)-based pose correction model [Zelinka and Kanis, 2020] is used to convert to 3D keypoints.

10 https://opus.nlpl.eu/OpenSubtitles/hi&en/v2024/OpenSubtitles

4.3 Data Processing

For each sign video, pose sequences are extracted and 51 upper-body keypoints are retained: 21 per hand, 2 eyes, 1 nose, 2 shoulders, 2 forearms, and 2 wrists. Missing hand detections can introduce redundant codes in Sign-Pose-VQVAE; this issue is mitigated by filling missing keypoints using spherical linear interpolation (Slerp). Keypoints are normalised using the left shoulder (kls), right shoulder (krs), and nose (kn) as {ki - kn - 3 × norm(kls, krs)}/{6 × norm(kls, krs)}, where ki denotes any keypoint.

4.4 EvaluationMetrics

Following prior SLP work [Wang et al., 2025; Tang et al., 2025b; Tang et al., 2025a; Saunders et al., 2020; Walsh et al., 2025], back-translation quality is evaluated using BLEU, WER, and ROUGE-L scores. For direct pose-level evaluation, we adapted standard metrics for sign language production: Dynamic Time Warping (DTW) handles temporal misalignment by optimally warping sequences, making it suitable for comparing pose motion performed at varying speeds, and Total Distance (TD) [Walsh et al., 2025] measures the ratio of overall distance the signer's dominant hand has moved in 3D space in reference and hypothesis to judge how expressive the productions are. Additionally, following Inan et al. [Inan et al., 2024], semantic similarity is measured using BERT-based precision, recall, and F1 scores (BERTP/R/F1) [Zhang et al., 2019] for both back-translation and gloss translation.

4.5 Ablation Study

For the text-to-gloss translation module, three strategies are explored: a rule-based method and two LLM-based models, GPT-4o-mini 12 (proprietary) andQwen3-4B 13 (open-source). All methods are applied to the SynICS-20k test set, and semantic similarity between input sentences and generated glosses is measured using BERT scores (Table 1). GPT-4omini achieves the best performance, but its gains over the rule-based approach are modest (0.0193, 0.0263, and 0.0228 in BERT-P/R/F1). Qwen3-4B shows similar marginal improvements. In contrast, the rule-based method is substantially more efficient, being approximately 12× faster than GPT-4o-mini and 2× faster than Qwen3-4B. Given its competitive performance and significantly lower computational cost, the rule-based approach is adopted for all subsequent experiments.

11 https://ai.google.dev/edge/mediapipe/solutions/vision/ holistic landmarker

12 https://platform.openai.com/docs/models/gpt-4o-mini

13 https://github.com/QwenLM/Qwen3

An extensive ablation study is conducted on the two main hyperparameters of Sign-Pose-VQ-VAE: codebook size and code dimension. Pose-to-Gloss translation is evaluated using the Sign Language Transformer [Camgoz et al., 2020]. Results are reported in Table 2. We compare a shared codebook (single stream across all keypoints) with separated codebooks (independent streams for body and hands). Separated codebooks consistently outperform the shared configuration across both WER and BLEU-1. The best performance is achieved with a codebook size of 1536 and a code dimension of 128, which is used in all experiments.

4.6 Quantitative Evaluation

The proposed method is compared against Progressive Transformers (PT) [Saunders et al., 2020] and the diffusion-based Sign-IDD [Tang et al., 2025b]. Both methods are trained on SynICS-20k with author-suggested optimal hyperparameter settings. Quantitative evaluation is performed via backtranslation using Sign Language Transformers (SLT) [Camgoz et al., 2020], trained on the synthetic dataset and evaluated on the linear synthetic test set, as well as on test sets generated by PT, Sign-IDD, and the proposed method. Results are summarised in Table 3. Across all metrics, the proposed approach significantly outperforms both baselines. Compared with PT, absolute improvements of 31.31 WER, 39.92

BLEU-4, and 0.0588 BERT-P are achieved. Relative to SignIDD, gains of 8.72 WER, 11.84 BLEU-4, and 0.0197 BERTP are obtained. Direct pose-level evaluation using DTW and TD on the dominant signing hand is shown in Table 3. The proposed method achieves the lowest DTW (0.94) and best TD (0.66), indicating superior temporal alignment and more expressive hand motion.

Additionally, the system is evaluated on the PHOENIX14T dataset using SLT back-translation, and results are reported in Table 4. Compared to the transformer-based PT [Saunders et al., 2020], the proposed method achieves 27.06 points WER reduction and 10.15 points BLEU-4 improvement on the validation set, and 26.21 points WER reduction and 9.44 points BLEU-4 improvement on the test set. Against graph-based methods (NAT-AT [Huang et al., 2021], NAT-EA [Huang et al., 2021]), the proposed method outperforms the strongest baseline, NAT-EA, by 9.86 points lower WER and 3.37 points higher BLEU-4 on the test set. Compared to diffusion-based methods (D3DP-sign [Shan et al., 2023], G2P-DDM [Xie et al., 2024], GDCM [Tang et al., 2025c], Sign-IDD [Tang et al., 2025b]), the proposed approach surpasses the best model, Sign-IDD, with 6.25 points WER reduction and 1.94 points BLEU-4 improvement on the validation set, and 4.51 points WER reduction and 0.95 points BLEU-4 improvement on the test set. Finally, compared to the state-of-the-art method LVMCN [Wang et al., 2025], the proposed method delivers 5.39 points lower WER and 1.70 points higher BLEU-4 on the validation set, and 3.28 points lower WER and 0.67 points higher BLEU-4 on the test set.

4.7 Qualitative Evaluation

The acceptability of the proposed SLP system in the Deaf community is assessed with three ISLRTC 14 certified ISL experts using 240 English sentences from a Hindi movie subtitles corpus. A web-based application was developed for this evaluation, where for each sample, the input text is presented alongside the generated avatar animation and a stick-figure visualization. Experts assign a score based on the rubrics: 1 - Not interpretable/minimally interpretable; 2 - Poor (some isolated words interpretable but overall intent unclear); 3 - Modest (overall intent clear but noticeable gestural errors and/or sign ordering error); 4 - Good (overall intent clear but some minor gestural errors); 5 - Excellent (fully interpretable with all glosses distinguishable). For samples rated below 5, experts additionally provide the glosses they interpreted from the animation, along with the expected glosses corresponding to the input text. The experts report an average rating of 4.33, with the rating distribution shown in Figure 4. Semantic similarity between the interpreted glosses (hypotheses) and expected glosses (references) is computed using BERT score, yielding BERT-P = 0.7683, BERT-R = 0.7497, and BERT-F1 = 0.7585.

Figure 3 compares the proposed approach with the HamNoSys-based avatar animation [Sugandhi et al., 2020] in ISL. The prior method resets the avatar to a rest pose after each sign (e.g., the first row, columns 2 and 4 in Figure 3), resulting in disjointed motion. In contrast, the proposed system generates temporally smooth transitions with effective co-articulation, resulting in more natural and continuous signing.

14 https://islrtc.nic.in/directory-of-isl-interpreters/

5 Conclusion

This work presents a scalable SLP system for low-resource settings that translates spoken or written language into 3D sign language avatar animations using isolated signs. The proposed Sign-Pose-VQ-VAE model learns discrete body poses and hand shapes while enabling smooth, co-articulated continuous sign generation. The system achieves state-of-theart performance on standard translation metrics for both Indian Sign Language (ISL) and German sign language. Overall, this approach overcomes key limitations of prior methods, particularly in fine-grained hand articulation, co-articulation quality, and scalability. The system is evaluated by ISLRTC certified ISL experts, receiving high ratings and positive feedback regarding its societal impact. The system with code is publicly available at https://cs.rkmvu.ac.in/∼isl/sl gen vqvae, aiming to strengthen accessibility and inclusion for the Deaf community, aligning with global goals for equitable communication.

Ethical Statement

This work uses the FDMSE-ISL [Patra et al., 2025b] and PHOENIX14T [Camgoz et al., 2018] datasets, which are publicly available. According to the authors, these datasets were collected with informed consent from the signers. We ensure participant confidentiality by using only pose sequences and avatar animations, without sharing any identifiable visual data.

Acknowledgments

We would like to thank Ms. Indira Ghosh for her assistance in creating the English sentence to ISL gloss corpus. We gratefully acknowledge the support from the Department of Science and Technology (DST), Government of India, through the DST-FIST grant [Sanction No.: SR/FST/MS-I/2022/116].